Deep learning-based solutions for precision agriculture have achieved promising results in recent times. Deep learning has been used to accurately classify different disease types and disease severity estimation as an initial stage for developing robust disease management systems. However, tracking the spread of diseases, identifying disease hot spots within cornfields, and notifying farmers using deep learning and UAS imagery remains a critical research gap. Therefore, in this study, high resolution, Unmanned Aerial System (UAS) acquired, Real-Time Kinematic (RTK) geotagged, RGB imagery at heights of 3 meters and 12 meters above ground level (AGL), was used to track disease hot spots in corn fields throughout the growing season. A total of 98,000 RGB images with a resolution of 8192 x 5460 pixels were acquired in cornfields located at Purdue University’s Agronomy Center for Research and Education (ACRE), using a DJI Matrice 300 with an RTK base station mounted with a 45-megapixel DJI Zenmuse P1 camera, from June 28th to August 31st, 2021. After carefully selecting images acquired at one-week intervals, they were split into multiple smaller tiles and super-pixels using the Simple Linear Iterative Clustering (SLIC) segmentation algorithm. Images were first split into tiles of sizes 250 x 250 pixels, 500 x 500 pixels, and 1000 x 1000 pixels, resulting in 726, 187, and 50 image tiles, respectively. Tiles that were not of 1:1 aspect ratio were padded with extra black pixels. Additionally, for SLIC segmentation, the images were split into the same number of super-pixels as the number of tiles using two different sigma values. For each tile and super-pixel, a 1:1 aspect ratio was maintained by placing the super-pixel on a black background for conducting a fair comparison with the tile approach. After the tiles and super-pixels were created, they were labeled as either soil, weed, healthy corn, or diseased corn. Five Convolutional Neural Network (CNN) architectures, namely VGG16, ResNet50, InceptionV4, DenseNet169, and Xception, were then used to train deep learning-based image classification models to compare the tile-based approach and SLIC segmentation for disease tracking in cornfields. After comparing the trained deep learning models using confusion matrices, testing accuracies, and F1 scores, the optimal number of tiles and SLIC segmented super-pixels were identified with the help of a plant pathologist. The deep learning model with the highest testing accuracy for disease tracking was then used to calculate the percentage of each diseased image by highlighting diseased regions. In addition, the RTK geolocation information for each image was used to update farmers with the location of disease hot spots within cornfields by developing a smartphone application and sending text message notifications.
